Class Activation Mapping and its variants are commonly used for explaining Convolutional Neural Network predictions.
Standard CAMs are vulnerable to adversarial manipulation like passive fooling, leading to misleading CAMs without affecting decision performance.
To address this vulnerability, Salience-Hoax Activation Maps (SHAMs) are introduced as a benchmark for CAM robustness under adversarial conditions.
DiffGradCAM, a novel approach to class activation mapping, is proposed to be resistant to passive fooling and matches standard CAM methods in non-adversarial cases.